Reproduce “BPEmb: Tokenization-free Pre-trained Subword Embeddings in 275 Languages” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Benjamin Heinzerling, Michael Strube
BPEmb provides pretrained byte-pair embeddings for 275 languages, training vectors for frequency-merged text pieces. Its pieces can be used without a separate tokenizer. In English mention-only entity typing, BPE-RNN scored 0.624 ± 0.029 versus 0.617 ± 0.007 for FastText-RNN; a 100,000-piece BPEmb configuration occupied 11 MB versus FastText’s 6 GB. Results vary across languages, and the evaluation excludes context.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.